Skip to content
SolvraONE

Knowledge

Human in the loop: when an AI agent hands over.

AI illustration: two rounded graphite blocks side by side; a channel of amber light leaves the left block and runs cleanly into a socket on the right block, like a handover
AI illustration
Published Fact-checked 14 Nov by a second AI (OpenAI) · 5 of 7 facts confirmed against the original source, 1 on site

In short · 7 min read

  • An AI agent should hand over to a person when it is unsure, the customer is upset, a money matter goes beyond a set approval limit, a legal question comes up or the customer explicitly asks for a person.
  • The handover is clean when the staff member sees the whole case at a glance and the customer does not have to repeat anything.

A good AI agent handles a lot on its own. A very good AI agent knows when it should not. Handing over to a person is not a failure of the agent but a fixed part of its work. What matters is that it happens by clear rules and that the person can then carry on without searching.

Five rules: when the agent hands over

Rule How the agent recognises it Why
Uncertainty The answer is not in the knowledge base, the data does not match or the question is ambiguous. Guessing costs more trust than an honest handover.
Emotion The customer is angry, desperate or writing about the same problem again and again. An upset person needs understanding and room to move that an agent does not have.
Money Refunds, credits, goodwill gestures, payment defaults or contract changes that go beyond a set limit. This is where real costs and liability arise.
Law Cancellation with a dispute, withdrawal, a complaint with a lawyer, a data protection request, a reference to damage. Mistakes have legal consequences; this belongs in trained hands.
Customer wants a person The customer asks for a person, a member of staff or “someone who can help me”. The customer’s wish comes first, without discussion.

The last rule is the most important. An agent that ignores the wish for a person or redirects it three times does more damage than it saves.

For decisions with legal effect, there is also a legal side: under Art. 22 GDPR, people in many cases have the right not to be subject to a decision based solely on automated processing, and may ask for a person to step in. This is not legal advice; clarify with your data protection adviser whether and how it applies to your case.

How a clean handover works

  1. Recognise: the agent notices that one of the five rules applies and stops deciding on its own.
  2. Announce: it tells the customer openly that their request is going to a person, and why, in one sentence; it only says a person is taking over right away if that is really the case.
  3. Summarise: it writes down for the staff member what the case is about, what has been checked and what is open.
  4. Route: the case goes to the right place, such as accounts, technical support or sales, with the right urgency.
  5. Set expectations: the customer learns what happens next and roughly when to expect an answer, only as precisely as you can keep to.
  6. Close: the staff member resolves the case and notes why it was handed over. This feeds back into the knowledge base.

What the customer should never experience: being cut off without explanation, an empty form, or being asked to tell everything again.

What the person sees at handover

The best handover is one after which the staff member does not have to ask. That is why these details belong in the case:

Detail Example
The request in one sentence “Customer wants a payment that was taken twice refunded.”
Reason for the handover Rule “money”: refund above the approved limit.
What has been checked Customer number confirmed, booking found in the system.
What is open Whether the second booking is an error or a separate order.
Mood calm, but in a hurry
Full history the whole conversation, not just the last sentence
The agent’s suggestion marked as a suggestion, never as a decision

The separation matters: what the agent has verified is shown differently from what it assumes. That way the staff member sees at once what they can rely on.

Measuring handovers

Handovers are one of the most important figures of an agent in operation. Four values make sense:

Figure The question behind it
Handover rate What share of conversations goes to people, and under which rule?
Handovers without reason How often could the agent have resolved the case itself? This can point to gaps in knowledge, unsuitable rules or technical limits; check the cause in the specific case.
Missed handovers How often should it have handed over and did not? These cases weigh most.
Time to takeover How long does the customer wait until a person replies?

A low handover rate is not automatically good. It can mean the agent decides too much on its own. So regularly check samples of the conversations it closed alone.

More on this

Large providers also separate this in their figures: Zendesk, for example, distinguishes cases the agent resolved itself from cases where it only prepared the work and a person resolved them, and only counts the first as a resolution. If you pay per resolved case, you should know exactly how handovers are counted. More on the page What does an AI agent in customer service cost?.

What the AI Act says

Article 14 of the AI Act

The EU AI Act requires human oversight for high-risk AI systems in Article 14 of the AI Act. A typical customer service agent often does not belong to this group. Independently of that, since 2 August 2026 the duty from Article 50(1) of the AI Act applies: in direct interaction, customers must be informed about the use of AI, unless this is already obvious from the circumstances and context. A clear handover to people is therefore not a duty for every agent, but good practice and the best answer to the question of who is responsible in the end. More on labelling on the page AI labelling under the AI Act. This is not legal advice.

Your next step

Read in full

Want handover rules that fit your business, or to check how your current chat hands over? We talk about it in a first call of 30 minutes: Get in touch.

What could AI take off your plate?

In a first call (30 minutes) we look at your work and tell you openly what is worth it and what is not.

Request the website check

Frequently asked questions.

When must an AI agent hand over to a person?

When it is unsure, when the customer is upset, when a money matter goes beyond a set approval limit, when a legal question comes up, and always when the customer asks for a person. These rules should be set down in writing and checked in operation.

Can a customer ask for a person at any time?

They should always be able to, without detours. For decisions with legal effect, there can also be a legal right to have a person step in.

What has to go along with the handover?

The request in one sentence, the reason for the handover, what has been checked and what is open, the customer's mood and the full conversation history. The agent's suggestions are marked as suggestions.

Is a high handover rate bad?

Not necessarily. What matters is whether the handovers were justified. Handovers without reason can point to gaps in knowledge, unsuitable rules or technical limits; missed handovers are the bigger risk.

Does the AI Act require a handover to people?

The AI Act requires human oversight for high-risk AI systems; a typical customer service agent often does not belong to this group. The duty to inform people about the AI interaction, on the other hand, applies to every AI system intended for direct interaction with people, unless the use of AI is obvious. This is not legal advice.

How quickly must a person take over?

There is no general deadline for a person to take over. Legal processing and response deadlines apply regardless, for example for requests for access or erasure: under Art. 12(3) GDPR, you provide information on the action taken without undue delay; the deadline is at most one month from receipt of the request. Where necessary because of the complexity or number of requests, it can be extended by two months; you inform the person of this within the first deadline and give the reasons. Only give the customer a waiting time your team can actually keep, and measure the time to takeover. This is not legal advice.

How is your website doing?

The check is free and without obligation. You get the report by email.

Request the website checkor ask for a first call

Sources.

  1. EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Art. 14 and Art. 50Vendor / primary · As of
  2. EUR-Lex: Regulation (EU) 2016/679 (GDPR), Art. 12 and Art. 22Vendor / primary · As of
All 3 sources
  1. Zendesk Help Center: About automated resolution tiersVendor / primary · As of

What does an AI agent cost in customer service?

What does an AI customer service agent cost? Pricing models in the market (per resolution, per seat, per conversation) and what drives cost, with sources.

AI labelling duty: what businesses have to label

AI labelling under the EU AI Act: what applies to chatbots, images, texts and social media, and how to put it into practice on your website, with sources.

AI and the GDPR: what the law asks of AI agents

AI agents and the GDPR: lawful basis, processing agreements, where data is processed and impact assessments. What businesses need to check.